Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

RDGCL: Reaction-Diffusion Graph Contrastive Learning for Recommendation

Jeongwhan Choi, Hyowon Wi, Chaejeong Lee, Sung-Bae Cho, Dongha Lee, Noseong Park

TL;DR

RDGCL presents a novel reaction-diffusion graph learning framework for recommendation, combining diffusion-driven low-pass filtering with reaction-driven high-pass filtering in a single-pass, cross-layer contrastive setup. By evolving initial embeddings through a RDG layer and contrasting diffusion and reaction views, it eliminates graph augmentations while enhancing accuracy and diversity. Empirical results across five real-world datasets show RDGCL achieving state-of-the-art performance and balanced recall with coverage and novelty, supported by ablations and robustness analyses. The approach offers a principled, efficient alternative to existing CL-based CF methods and opens avenues for further optimization of reaction dynamics in graph-based learning.

Abstract

Contrastive learning (CL) has emerged as a promising technique for improving recommender systems, addressing the challenge of data sparsity by using self-supervised signals from raw data. Integration of CL with graph convolutional network (GCN)-based collaborative filterings (CFs) has been explored in recommender systems. However, current CL-based recommendation models heavily rely on low-pass filters and graph augmentations. In this paper, inspired by the reaction-diffusion equation, we propose a novel CL method for recommender systems called the reaction-diffusion graph contrastive learning model (RDGCL). We design our own GCN for CF based on the equations of diffusion, i.e., low-pass filter, and reaction, i.e., high-pass filter. Our proposed CL-based training occurs between reaction and diffusion-based embeddings, so there is no need for graph augmentations. Experimental evaluation on 5 benchmark datasets demonstrates that our proposed method outperforms state-of-the-art CL-based recommendation models. By enhancing recommendation accuracy and diversity, our method brings an advancement in CL for recommender systems.

RDGCL: Reaction-Diffusion Graph Contrastive Learning for Recommendation

TL;DR

RDGCL presents a novel reaction-diffusion graph learning framework for recommendation, combining diffusion-driven low-pass filtering with reaction-driven high-pass filtering in a single-pass, cross-layer contrastive setup. By evolving initial embeddings through a RDG layer and contrasting diffusion and reaction views, it eliminates graph augmentations while enhancing accuracy and diversity. Empirical results across five real-world datasets show RDGCL achieving state-of-the-art performance and balanced recall with coverage and novelty, supported by ablations and robustness analyses. The approach offers a principled, efficient alternative to existing CL-based CF methods and opens avenues for further optimization of reaction dynamics in graph-based learning.

Abstract

Contrastive learning (CL) has emerged as a promising technique for improving recommender systems, addressing the challenge of data sparsity by using self-supervised signals from raw data. Integration of CL with graph convolutional network (GCN)-based collaborative filterings (CFs) has been explored in recommender systems. However, current CL-based recommendation models heavily rely on low-pass filters and graph augmentations. In this paper, inspired by the reaction-diffusion equation, we propose a novel CL method for recommender systems called the reaction-diffusion graph contrastive learning model (RDGCL). We design our own GCN for CF based on the equations of diffusion, i.e., low-pass filter, and reaction, i.e., high-pass filter. Our proposed CL-based training occurs between reaction and diffusion-based embeddings, so there is no need for graph augmentations. Experimental evaluation on 5 benchmark datasets demonstrates that our proposed method outperforms state-of-the-art CL-based recommendation models. By enhancing recommendation accuracy and diversity, our method brings an advancement in CL for recommender systems.
Paper Structure (40 sections, 1 theorem, 16 equations, 12 figures, 9 tables)

This paper contains 40 sections, 1 theorem, 16 equations, 12 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Preliminaries & Related Work
  3. Graph Filters and GCN-based CFs
  4. Contrastive Learning for Recommendation
  5. Reaction-Diffusion Equations
  6. Neural Ordinary Differential Equations
  7. Proposed Method
  8. Overall Architecture
  9. Reaction-Diffusion Graph (RDG) Layer
  10. Diffusion Process as Low-pass Filter
  11. Reaction Process as High-pass Filter
  12. Cross-layer Contrastive Learning
  13. Model Complexity
  14. Property of RDG Layer
  15. Relation to Other Methods
  16. ...and 25 more sections

Key Result

Theorem 1

The RDG layer applies a $2\tilde{\mathbf{A}} - \tilde{\mathbf{A}}^2$ filter, which emphasizes mid-to-high frequency components of the graph signal more than the $\tilde{\mathbf{A}}$ filter.

Figures (12)

  • Figure 1: Comparison in terms of Recall@20 and $h_{\text{RN}}$@20, the harmonic mean of the recall and novelty (see Sec. \ref{['sec:trade']}), on Yelp. Our RDGCL outperforms SGL Wu2021SGL, SimGCL xu2023simdcl, and LightGCL cai2023lightgcl.
  • Figure 2: The architectures of SimGCL, LightGCL, XSimGCL, and RDGCL.
  • Figure 3: The illustration of our RDGCL where we solve our reaction-diffusion system with the Euler method. LPF (resp. HPF) stands for the low (resp. high)-pass filter.
  • Figure 4: Trade-off among the three metrics
  • Figure 5: Performance on users of different sparsity degrees
  • ...and 7 more figures

Theorems & Definitions (2)

  • Theorem 1
  • proof